Nowadays, online social media platforms enable individuals or groups to freely express their opinions. Sometimes such opinions lead to personal attacks, harassment, and abusive language. In natural language processing, detecting and classifying toxicity is one of the key challenge. Therefore, we first focus on subsymbolic artificial intelligence (AI) for multi-label toxicity identification, where an instance is mapped to a set of pre-defined label categories. Second, we propose a hybrid AI to analyze only negative toxic comments. The hybrid AI is the combination of symbolic AI and subsymbolic AI. First, the symbolic AI can be used to measures three sentiment polarities (positive, negative, and neutral) of the given text. Second, the subsymbolic AI uses only the negative sentiment polarity texts to map a set of pre-defined label categories. In particular, we use the BERT fine-tuning model as subsymbolic AI for multi-label toxicity category detection. This proposed approach is employed on the Wikipedia’s talk page toxic comment classification dataset. It includes 159,571 instances as base dataset and 223,549 instances as large dataset with six label categories for the task of multi-label classification. The example, label, and rank-based multi-label metrics are used to evaluate the performance of the proposed hybrid AI model. Results indicate that the subsymbolic-based BERT model outperforms with both base (98.61%) and negative reviews base data (97.80%) than the large dataset.